m9 and rng polishing

zephcore/boards/common/esp32_common.conf

@@ -60,8 +60,30 @@ CONFIG_BT_BUF_EVT_RX_COUNT=14
 CONFIG_ESP32_BT_CTLR_LE_SECURITY_ENABLE=y
 
 # ========== Heap ==========
-# ESP32 BLE stack requires larger heap (override zephcore_common 2KB default)
-CONFIG_HEAP_MEM_POOL_SIZE=32768
+# ESP32 BLE stack requires larger heap (override zephcore_common 2KB default).
+#
+# 32768 was NOT enough and failed non-obviously.  The Espressif controller
+# allocates its exchange memory through esp_bt_malloc_func(), which the Zephyr
+# port maps to k_malloc() (esp_heap_adapter.h, CONFIG_ESP_BT_HEAP_SYSTEM=y) —
+# i.e. straight out of this pool.  esp_bt_controller_init() asks for 0x7800 =
+# 30720 bytes in ONE allocation: 93.75% of a 32 KB pool, leaving 2 KB for block
+# overhead and every other k_malloc user that ran before bt_enable().  Boards
+# that allocated little before BLE init happened to fit; ThinkNode M9 (display +
+# colour overlay, STC8H keypad, LR1110) did not, and the blob asserted inside
+# its own allocator instead of returning ESP_ERR_NO_MEM:
+#
+#   BLE assert emi.c 164, param 00000000 00007800   <- (NULL, requested size)
+#
+# Zephyr's own driver anticipates this and prints "Consider increasing
+# CONFIG_HEAP_MEM_POOL_SIZE" (drivers/bluetooth/hci/hci_esp32.c) — but only on
+# the graceful path the blob does not take.  Every ESP32 board was one
+# allocation away from the same failure, so this is raised platform-wide.
+#
+# Headroom verified by build on all 12 shipped ESP32 boards: the S3/C3/C6 boards
+# have 66-135 KB free internal DRAM and absorb this easily.  ttgo_tbeam is the
+# sole exception (99.45% DRAM, 768 B free) and overrides it back in its
+# board.conf — see the note there.
+CONFIG_HEAP_MEM_POOL_SIZE=65536
 
 # ========== Flash ==========
 # 4MB is standard for XIAO/LilyGo RISC-V boards.

zephcore/boards/esp32/thinknode_m9/thinknode_m9_esp32s3_procpu.dts

@@ -47,13 +47,26 @@
 	};
 };
 
-/* Reduce CPU clock from 240 MHz to 160 MHz — HAL Kconfig minimum supported */
+/* 240 MHz — matches every other ESP32-S3 board in the tree (xiao_esp32s3,
+ * station_g2, heltec_*).  This was 160 MHz with the comment "HAL Kconfig
+ * minimum supported", which is not correct: hal/espressif/zephyr/Kconfig
+ * derives ESP_DEFAULT_CPU_FREQ_MHZ straight from this DT property and offers
+ * 96/120/160/240 with no floor at 160.  Nothing on this board requires the
+ * lower clock (the octal PSRAM runs at 40 MHz off its own divider, and
+ * station_g2 pairs the same OPI PSRAM with 240 MHz).
+ *
+ * Kept deliberate rather than "faster is better": the node is event-driven and
+ * idles in k_event_wait almost always, so race-to-idle makes the clock a weak
+ * lever on battery life, while the 320x240 colour panel is the one genuinely
+ * CPU-bound consumer.  240 also removes this board's only timing anomaly,
+ * which matters while the CPU-jitter entropy source is under investigation
+ * (jitter health check failed here — see memory/findings.md). */
 &cpu0 {
-	clock-frequency = <DT_FREQ_M(160)>;
+	clock-frequency = <DT_FREQ_M(240)>;
 };
 
 &cpu1 {
-	clock-frequency = <DT_FREQ_M(160)>;
+	clock-frequency = <DT_FREQ_M(240)>;
 };
 
 &usb_serial {

zephcore/boards/esp32/ttgo_tbeam/board.conf

@@ -41,3 +41,17 @@ CONFIG_ESPTOOLPY_FLASHMODE_DIO=y
 CONFIG_ZEPHCORE_MAX_CONTACTS=140
 CONFIG_ZEPHCORE_MAX_CHANNELS=8
 CONFIG_ZEPHCORE_OFFLINE_QUEUE_SIZE=128
+
+# BLE controller heap — hold this board at the old 32 KB.
+#
+# esp32_common.conf raises CONFIG_HEAP_MEM_POOL_SIZE to 65536 because the
+# Espressif controller allocates 30720 bytes out of it in one go and 32 KB left
+# no margin (see the long note there).  This board cannot pay for it: measured
+# at 139684/140452 B = 99.45% DRAM, only 768 bytes free, so +32 KB does not
+# link.  Holding it here keeps T-Beam exactly as it builds today.
+#
+# NOT a fix — this board still runs the controller on the same 2 KB margin that
+# broke ThinkNode M9, and its classic-ESP32 controller draws from the same pool.
+# Making T-Beam safe needs DRAM reclaimed (contacts/queue are already trimmed
+# above), not a larger number here.  Tracked separately.
+CONFIG_HEAP_MEM_POOL_SIZE=32768